Writer Identification Using Siamese Networks and Character-Level Analysis

Writer Identification Using Siamese Networks and Character-Level Analysis

Stať ve sborníku v databázi WoS či Scopus

Writer identification of handwritten text is a task in forensic document analysis, traditionally relying on visual comparison by experts. However, this process is time-consuming and subjective. Although the amount of crime involving handwriting has remained constant, the overall volume of handwritten material has decreased. This paper presents an approach based on a Siamese Neural Network (SNN) to writer identification by analyzing a limited information source - just a single character – A – from samples of Czech handwriting. The main contributions are: (i) the design of an SNN with a five-block convolutional branch combined with voting strategies incorporating an uncertainty zone; and (ii) a detailed experimental comparison of over 400 architecture and hyperparameter configurations in terms of accuracy, F1-score, and decision efficiency. The best model achieved relatively high accuracy – 96.1 % accuracy and a F1-score of 0.932 while abstaining from classification in approximately 57% of ambiguous cases. The trade-off between classification confidence and coverage, the limitations of single-character analysis, and the potential for generalization to open-set scenarios and multimodal inputs are discussed. The proposed approach offers an objective and reproducible method suitable for forensic handwriting analysis.

Writer identification of handwritten text is a task in forensic document analysis, traditionally relying on visual comparison by experts. However, this process is time-consuming and subjective. Although the amount of crime involving handwriting has remained constant, the overall volume of handwritten material has decreased. This paper presents an approach based on a Siamese Neural Network (SNN) to writer identification by analyzing a limited information source - just a single character – A – from samples of Czech handwriting. The main contributions are: (i) the design of an SNN with a five-block convolutional branch combined with voting strategies incorporating an uncertainty zone; and (ii) a detailed experimental comparison of over 400 architecture and hyperparameter configurations in terms of accuracy, F1-score, and decision efficiency. The best model achieved relatively high accuracy – 96.1 % accuracy and a F1-score of 0.932 while abstaining from classification in approximately 57% of ambiguous cases. The trade-off between classification confidence and coverage, the limitations of single-character analysis, and the potential for generalization to open-set scenarios and multimodal inputs are discussed. The proposed approach offers an objective and reproducible method suitable for forensic handwriting analysis.

writer identification, handwriting text, forensic analysis, Siamese network, uncertainty modeling

writer identification, handwriting text, forensic analysis, Siamese network, uncertainty modeling

MYŠKA, V.; BURGET, R.; JEŽEK, Š.; LUŇÁKOVÁ, M.; MORAVCOVÁ, P.; DOUBKOVÁ, Š.; MEZINA, A.; JONÁK, M.

2026

03.12.2025

IEEE

Italy

979-8-3315-7675-2

2025 17th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT)

182

186

5